Detailed structural information is essential for understanding biological processes and for providing the basis for the development of therapeutic strategies against a variety of diseases. Experimental methods allow the accurate determination of high-resolution structures, but they are encumbered by significant effort and experimental constraints. As an alternative, modern computational methods can predict protein structures to a good degree of accuracy. However, computational models often do not reach experimental accuracy. Computational structure refinement aims at improving initial models towards experimental quality. Successful protein structure refinement requires sampling methods that can generate native-like conformations and scoring methods that are able to identify the most native structures from a set of candidates without knowledge of the true experimental structure. Building on past progress with the development of structure refinement methods, new methods are developed to further improve the quality and efficiency of sampling and the selection of native-like structures. A key component is the use of the intermediate resolution model PRIMO in conjunction with fully atomistic models. These methods will be applied to the refinement of both soluble and membrane-bound proteins. The development of a practical protocol that can be made available to the community via a web service is a central goal of this proposal.